犬种识别CNN模型训练：如何将Stanford Dogs Dataset重构为Train/Test按犬种分类的文件夹结构

Alright, let's fix this folder structure for your ResNet training—Stanford Dogs Dataset actually comes with official train/test split files, so we can automate this instead of doing it manually. Here's exactly what to do in your Colab notebook:

Step 1: Confirm Your Raw Dataset Structure #

First, make sure your extracted dataset has these key components in its root folder (let's assume it's stored at /content/stanford_dogs):

• Images/: Contains subfolders named like n02085620-Chihuahua, each holding images for that specific breed.
• train_list.txt: Official list of training images (each line includes the relative path to an image and a label).
• test_list.txt: Official list of test images, formatted the same way as the training list.

Step 2: Run the Python Script to Restructure #

Copy this code into a Colab cell, adjust the paths if your dataset is stored elsewhere, and execute it:

import os
import shutil

# --------------------------
# Update these paths if needed
# --------------------------
base_raw_dir = "/content/stanford_dogs"  # Path to your extracted dataset
target_split_dir = "/content/stanford_dogs_split"  # Where the structured dataset will be saved

# --------------------------
# Set up source and destination paths
# --------------------------
raw_images_dir = os.path.join(base_raw_dir, "Images")
train_split_file = os.path.join(base_raw_dir, "train_list.txt")
test_split_file = os.path.join(base_raw_dir, "test_list.txt")

train_dest_dir = os.path.join(target_split_dir, "train")
test_dest_dir = os.path.join(target_split_dir, "test")

# Create base folders if they don't exist
os.makedirs(train_dest_dir, exist_ok=True)
os.makedirs(test_dest_dir, exist_ok=True)

# --------------------------
# Function to split and copy images
# --------------------------
def organize_images(split_file_path, destination_dir):
    with open(split_file_path, "r") as file:
        all_lines = file.readlines()
    
    for line in all_lines:
        # Split line to get the relative image path (ignore the label at the end)
        image_rel_path = line.strip().split()[0]
        breed_folder_name, image_filename = image_rel_path.split("/")
        
        # Build source and destination paths
        source_image_path = os.path.join(raw_images_dir, breed_folder_name, image_filename)
        breed_dest_folder = os.path.join(destination_dir, breed_folder_name)
        
        # Create breed folder in target if it doesn't exist
        os.makedirs(breed_dest_folder, exist_ok=True)
        dest_image_path = os.path.join(breed_dest_folder, image_filename)
        
        # Copy the image to the target location
        shutil.copy(source_image_path, dest_image_path)

# --------------------------
# Run the organization for train and test sets
# --------------------------
print("Organizing training set...")
organize_images(train_split_file, train_dest_dir)

print("Organizing test set...")
organize_images(test_split_file, test_dest_dir)

print("Done! Check the structured dataset at:", target_split_dir)

Step 3: Verify the Result #

After running the script, navigate to /content/stanford_dogs_split in Colab's file explorer. You'll see:

• train/: Contains subfolders for each dog breed, with all corresponding training images inside.
• test/: Mirror structure, but populated with test images.

This structure is perfect for using with Keras' ImageDataGenerator or PyTorch's ImageFolder—both tools will automatically map each breed folder to a unique class label, which is exactly what you need for your ResNet training pipeline.

A quick heads-up: The dataset has ~20k images total, so the copy process might take a minute or two depending on Colab's current speed. No need to panic if it doesn't finish instantly!

内容的提问来源于stack exchange，提问作者Ayushya Pare